Recently, I've been working on a B2B agent. One of the business modules involves task creation. I wanted the agent to be able to supplement missing fields through memory and simple dialogue before task establishment, completing task clarification so that the agent could execute tasks with clear objectives.
When orchestrating this module, I made a fatal mistake: I preset the "product manager" role's requirement clarification methodology in the prompt. This caused the agent to rigidly follow all preset procedures step-by-step regardless of what questions I asked, even when I had already provided all the key elements.
So I had to stop and use GPT to redesign the prompt. Unexpectedly, GPT reminded me that the issue wasn't the "product manager" role design, but rather that I had designed the workflow as a process-driven model.
This was both heartbreaking and ironic, because I had always opposed simply characterizing AI with roles and strongly disliked making Agents into workflow engine shells. Yet when it came to building products myself, I still fell back into familiar work patterns: defining roles, designing SOPs, and then letting AI execute step-by-step.
Many people understand workflow as step-by-step execution, letting AI perform actions sequentially, like Dify's traditional workflow orchestration tools. However, since last year when Codex, Claude Code, and other coding agents debuted, the definition of workflow has already shifted from step-by-step to loops centered around goals. What we need to do is define the goal, then let AI revolve around that goal, continuously looping within harness constraints and capability support until the goal is achieved.
In plain terms: tell AI what you want, what support you can provide for AI, what AI cannot do, and how to handle problems when they arise—then let AI freely handle the rest. This is what constitutes an agent's workflow.
Traditional workflow orchestration has value in terms of certainty and cost-effectiveness, but an agent's core capability is dynamically selecting paths around goals, enabling it to approximate results within dynamic data spaces. Designing an agent with workflow orchestration thinking is like putting a V8 engine on an electric scooter—it can run, but it's completely cost-ineffective.
Moreover, this approach transfers complexity that should be borne by AI back onto the user. Meanwhile, due to LLM hallucination issues, monitoring and verification costs rise sharply.
The workflow of a true AI agent should be "step by step to goal" or rather "goal to GOAL." Humans define the task's ultimate goal, boundaries, resources, risks, and acceptance criteria; AI is responsible for breaking down steps around this goal, calling tools, processing intermediate results, and continuously adjusting paths during execution.
Therefore, the focus of workflow is not designing every step in advance, but rather excavating requirements to complete goal definition, then handing over step generation and execution to AI.